Yintoni kunokwenzeka xa uxhumane i-AI yakho kwiindawo ezifanelekileyo

Le nqaku elandelayo: bheka post yokuqala ke.

I-AI Principle II: Load Prompts Safely (Uma ufuna ukuba)

Ufuna i-chatbot yakho ukuqala ukuxhumana neengxelo ze-Taylor Swift emzimbeni yokufikelela i-tech support? Yinto yaye i-chatbot yethu yenzelwe xa sinxulumene neempawu ezidlulileyo. Ukuba ufuna ukunciphisa isicelo yakho kwaye ukwenza i-AI ye-architecture yakho emangalisayo, nceda uqhagamshelane. (Sorry Taylor abalandeli!)

Where to Store Imibuzo

Ukuhluza iingcebiso zakho kunye neengcebiso ezininzi? Okanye ukulayisha kwizithuthi ezininzi? Mhlawumbi iingcebiso ezimbini? Ngezantsi i-framework yokufunda le nqakraza.

I-Option A - I-Store Prompts kwi-Git

Ixabiso yokuqala kufuneka uqhagamshelane: Ngaba kukho isizathu esizodwa sokugcina i-prompts ngaphandle kwe-code yakho? Ukuba akukho, nceda uqhagamshelane i-prompts kwi-Git kunye ne-codebase, apho ziquka. Oku kunokuba yinkonzo elula kakhulu kwaye lula ukuhambisa. Kuyinto ingqalelo olungaphakathi.

Wagqiba kwi-Principle #1: I-Prompts yi-Code. Ukugcina iindawo ze-codebase yakho ngaphandle kwe-Git kunokwenzeka kwaye ngexesha kufuneka, kodwa ayikho izinto ezincinane. Unxibelelwano yokubhalisa i-Prompts ngokufanelekileyo.

I-Option B - I-Load Prompts ukusuka kwi-Version-Controlled Platform

Yintoni ukuba ezinye iingcebiso zakho ziquka kubhalwe yi-non-Engineers? Oku kunokwenzeka ukuba iingcebiso ye-domain ye-deep kufuneka kufuneka. Okanye ukuba iingcebiso kufuneka ifunyenwe ngexesha elide kwaye unako ukufikelela kwi-engineering department.

Ukuba oku, kufuneka ukulanda i-prompt ngexesha lokugqibela ukusuka kwi-version-controlled source. Ndiyathanda i-Confluence kunye ne-Google Docs ezisetyenziswa ngempumelelo ekusebenziseni oku. Ezinye i-version-controlled, i-API-access platform ziyafumaneka.

Ukuba ukutshintsha i-logic yokuthumela okushicilela, musa ukunciphisa ubunzima ukongeza le nqakraza. Uya kufuneka uqhagamshelane nezimo ezininzi zeengxaki kunye neengxaki ukuze ufumane isicelo yakho. Iintlawulo ze-access kufuneka ifakwe kunye nokugcina, kwaye ukulawula okuzenzakalelayo kunye neengxaki ezongezelelweyo kufuneka ifakwe iingxaki ngokukhawuleza kunokwenzeka.

Indawo ezininzi iingxaki ukuba kufuneka uqhagamshelane:

• I-application is not able to load prompts at runtime. Do you kill the deployment? Switch to a backup version of the prompt?
• I-Prompt syntax ifumaneka emva kokuguqulwa kwaye ivumela izakhiwo zeendaba ezisetyenziswayo. Ukuhlolwa okuzenzakalelayo awukwazi ukufumana ingxaki ngenxa yokuba ama-prompts abaxhasiwa ngexesha lokuvelisa. Yintoni i-infrastructure ye-test kunye ne-monitoring ezongezelelweyo kufuneka ifumaneke ukuze ifumanise oku kwaye ukunciphise imiphumo yeenkcukacha?
• I-Prompt kufuneka ilawulwe ngempumelelo emva kokuguqulwa. Yintoni oku kufuneka ifumaneke
• I-application is not able to load prompts at runtime. Ngaba ushiye ukuqhagamshelwano? Qhagamshelane kwi-backup version of the prompt?
• I-Prompt syntax ifumaneka emva kokuguqulwa kunye nokuguqulwa kwizakhiwo zeendaba ezisetyenziswa. Ukuhlolwa okuzenzakalelayo awukwazi ukucacisa inkinga ngenxa yokuba i-prompts ayidlalwa ngexesha lokusebenza kwe-test. Yintoni i-infrastructure ye-test kunye ne-monitoring eyongezelelweyo ukucacisa oku kwaye ukunciphisa imiphumo ye-client?
• I-Prompt kufuneka ifumaneke ngempumelelo. Yintoni kufuneka i-deployment ye-code entsha? Oko ukwakha i-UI eyahlukileyo ye-deployment ye-prompt?
• I-Syntax eyongezwa kwi-document ngu-Platforms efana ne-Confluence inokufumanisa i-runtime prompt, okufanayo ukusebenza yayo. Qinisekisa ukuba uthathe i-fuzz kwi-tools efana ne-Beautiful Soup.




Zonke iingxaki ziye 100% zibonakalayo. Kodwa kulula ukufikelela ukuba ukulayisha i-prompt ukusuka kwi-Google Doc kuyinto isebenzo esincinane leyo akayi kucacisa i-architecture ye-application ngokubanzi. Njengoko iye yaziwa apha, ukulayisha i-prompt ye-external kuyinto ibhizinisi elidlulileyo ukuba ifumaneka ngokufanelekileyo kwi-applications eziqhelekileyo.

Option C - Load Prompts ukusuka Non-Version-Controled Platform

I-idea ebuthayo, kwaye uya kufuna. Umgangatho we-truth yeengxaki kufuneka i-version-controlled, i-API efanelekileyo kunye ne-access controls. Oku akuyona indawo yokunciphisa iingongoma.

I-Option D - I-Hybrid Approach

I-hybrid approach ibandakanya ukugcinwa kwezinye amaxwebhu ngqo kwi-codebase yakho kunye nokukhutshwa kwamanye amaxwebhu e-external, i-version-controlled sources. Nangona ukugcina indawo epheleleyo kumadokhumenti bonke kubandakanya kubandakanya kubandakanya kakhulu, kukho iimeko apho isicwangciso se-hybrid inokukwazi ukunika izinzuzo.




Ukuhlola ukulungiswa kwe-hybrid under conditions such as:

• Ukusetyenziswa okuxhaswa: Izimpendulo ezininzi ziquka iimveliso ezininzi ze-domain non-coding, okuvumela ukulayisha kwebhizinisi, kwaye ezininzi ziya kubhalwe kuphela ngu-engineers.
• Ukuhlolwa kweRisiko: Izimpendulo ezininzi (isib. i-guardrails) kufuneka ziyafumaneka kwi-repository esemthethweni ukuze kube lula kakhulu. Izimpendulo ezininzi ezininzi, ikakhulukazi izimpendulo ezininzi ezincinane, ziyafumaneka ngokufanelekileyo kwi-external.
• I-Evaluation Flexibility Ukusetyenziswa kwe-Mixed: Ezinye amaxwebhu zihlangene izibuyekezo ezininzi ze-domain non-coding experts, kwenza ukulayisha kwe-external ngokufanelekileyo, kwaye ezinye ziya kubhalwe kuphela ngu-engineers.
Ukusetyenziswa kwe-Mixed
• Ukulawulwa kweRisk: Iingcebiso zokuzonwabisa (isib. iingcebiso zokuzonwabisa) kufuneka ziyafumaneka kwi-repository yokuqala ukuze kube lula kakhulu. Iingcebiso zokuzonwabisa ezincinane, ikakhulu iingcebiso zokuzonwabisa ezininzi, ziyafumaneka ngokhuseleko kwi-external.
Ukulawulwa kweRisk
• I-Evaluation Flexibility: I-Prompts eyenziwe kwi-evaluation ye-ML-style ingasebenza kwi-externally ukuze kubuyele ukuhlanganiswa kwinkqubo ye-evaluation.
Ukuhlaziywa Kwakhona

I-Guardrail Prompts

I-Guardrail prompts (eyaziwa nangokuthi i-censor prompts) zihlanganisa imibuzo ngaphambi kokufika kubasebenzisi, ukuphucula izisombululo ezifanelekileyo, ezigqongileyo, kunye ne-compliant. I-Guardrail isebenza njenge-mecanism yokhuseleko, ikakhulukazi kwizicelo apho izisombululo ze-usebenzisa zihlanganisa iingxaki ezininzi ze-legal okanye ze-ethical. Zibonisa i-second line of defence, ukufumana izisombululo ze-inappropriate that slip through.




Qhagamshelana neengxaki ze-guardrail ezivela kwi-doc ze-external - oku kwandisa i-risk engabikhoyo. Okanye ukugcina kwi-Git kunye ne-code yakho okanye usebenzisa isixhobo se-third-party ekhethekileyo, ezifana ne-I-Fiddle Guardrails. I-Guardrail isisombululo akufutshane kalula kakhulu, ngoko le mgangatho akufutshane kakhulu.

I-Fiddle Guardrails




Ukusetyenziswa kwe-guardrails iyisiseko yayo, kuya kuxhomekeke ngakumbi kwi-post elandelayo. It is a great pattern that improves the safety of your application and helps you sleep better at night. Just don't load them from Google Docs.

I-Loading Prompts for Easier Evaluation

Iqela ziquka i-prompts kwi-externally ukuze zihlanganise iinjini ze-evaluation, ezifana ne i-ML Flow. I-assumption esisiseko esekelwe kwinkqubo ye-prompts yinto efana neemodeli ze-ML kwaye kufuneka ukuhlaziywa kwimodeli ze-statistics. Unxibelelanisa i-prompt, ukunyaniseka i-F1 iqhosha kwi-output (okanye i-metric efanelekileyo) kwaye uqhagamshelane.

I-ML Flow




Le nqakraza kunokwenzeka ngamanye amaxesha—umzekelo, kwi-clasification prompts eyenziwe ukuba ifumaneke njengama-ML models. Kodwa i-prompts ezininzi ziquka kwi-Principle #1: I-LLM Prompts Are Code. I-prompts ezininzi ziyafumaneka kwi-application logic kunokuba kwi-ML models. Zifumaneka ngakumbi ukuhlaziywa kwe-Pass-Fail kunye ne-code efanayo, kunokuba yi-statistical evaluation approach.

I-LLM Prompts Are Code




Iinjini ze-evaluation ze-external ayikwazi ukunceda kunye neengxaki ezininzi. Ngoku, kufuneka usebenzise i-AI-driven tests, efana neengxaki ze-unity ezivamile. Ezi ziya kubalulekile kwiiposi ezilandelayo.




Qaphela iimeko ezilandelayo:

• Izixhobo ezininzi ezisebenzayo ezibonakalayo iimodeli ze-machine learning (isib. iingxaki ze-classification okanye iingxaki ze-scoring) kufuneka ziye ziye ziye ziye ziye ziye zibonakalayo.
• Izixhobo ezininzi ze-business logic ziquka kwi-codebase yokuqala, ezisetyenziswa kwiinkqubo ezijwayelekile ze-automated testing efana neenkqubo ze-unity testing.
• Where evaluation external is warranted, isolate only those prompts, where possible.
• Iintloko kuphela ezifuneko eziquka ngokucacileyo iimodeli ze-machine learning (isib. iingxaki ze-classification okanye iingxaki ze-scoring) kufuneka ziye ziye ziye zibonwa.
• Izixhobo ezininzi zeengcali zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi zebhizinisi.
• Ukuba i-evaluation ye-external iyafumaneka, i-isolate kuphela amaxwebhu, xa kunokwenzeka.

Ukuhlolwa Kwimeko

I-central issue kunye ne-loading prompts yi-availability - yintoni kufuneka uyenze ukuba i-prompt ayibhalwe xa ufuna.




Ezi yinto yaye yaye kwakhona kwimeko ye-Taylor Swift. Akukho yeemvakalelo ye-app ye-tech support efakwe ngenxa yeemvakalelo ye-Confluence, kuquka iimvakalelo ye-guardrail. Oku ngoko ke akuyona iimvakalelo ze-runtime kwaye i-bot uqala ukuxhaswa ngaphandle kwezinye imiyalelo okanye iingxaki (kuba i-input formating string iye yindawo yeemvakalelo). Yaye yintoni i-OpenAI's LLM ufuna ukuxhaswa ngaphandle kweemvakalelo? Kuboniswa - iingxelo ze-"I Want to Break Free" ye-Queen kunye neengxaki ezininzi zeTaylor Swift. Ngafanelekileyo, oku lithathwe kwaye




Yintoni iingxaki ezininzi ziyafumaneka? Iingxaki ezimbini ziyafumaneka:

• Khumbula ukuba amaxwebhu ezisetyenziswa ngempumelelo. Kukho ingxaki eyenziwe ngexesha lexwebhu ebonakalayo, njengoko le app ayikwazi ukusebenza ngaphandle kwexwebhu ebonakalayo.
• Ixwebhu yexwebhu ebonakalayo ebonakalayo kwi-externally kunye nexwebhu ezininzi. Yinto yexwebhu ebonakalayo ebonakalayo ngexwebhu ngexesha elifanelekileyo. Oku kufuneka ibekwe kwi-Git njenge-line lokugqiba lokugqiba.
• Akukho ukulungiselela ukuba amaxwebhu zithunyelwe ngempumelelo. Kukho ingxaki eyenziwe ngexesha lexwebhu zithunyelwe ngempumelelo, njengoko le app ayikwazi ukusebenza ngaphandle kwexwebhu zithunyelwe.
• I-guardrail prompt iye ifakwe kwi-externally kunye ne-rempts ezininzi. Yintoni i-prompt ebangelwa ifakwe ngexesha elifanayo. Yena kufuneka ifakwe kwi-Git njenge-last line ye-defense.




Emva kwimeko, i-guardrail prompt yaye-migrated kwi-Git kwaye i-exceptional logic iyathunyelwa ukucacisa ukuhanjiswa ukuba i-prompt ayikwazanga ukuyisebenzisa okanye ayikwazanga. Uyakwazi ukucacisa kwi-postmortem ngokufanelekileyo ngokufanayo nezi zixazululo.

Ukuhlaziywa

Kwi post, ndandisa iingxaki ezininzi malunga nokupakisha kunye nokupakisha kwizicelo ze-AI. Isiseko se-default i-storage ye-prompts kunye ne-code yakho kwi-repository ye-version-controlled. Kwixesha nje xa kukho isizathu esibonakalayo, njenge-editing embalwa ye-non-engineers okanye izicelo ze-evaluation ezizodwa.




Ukuba i-prompts kufuneka ifakwe kwi-externally, ukhethe iiyunithi ezifanelekileyo, kunye ne-version-controlled, ukongeza ukucacisa kunye nokulawula ukhuseleko. I-Guardrail prompts, ngokutsho umxholo wayo yokhuseleko kwe-application, kufuneka ibekwe kwi-codebase yakho ukuze ukunceda iingxaki ezininzi zokhuseleko.




Izixhobo ezininzi ziquka ngokwemvelo kwi-code kunokuba kwiimodeli ze-ML, ngoko usebenzisa i-ML style tools kuphela apho unayo unemfuneko. Ungabhalisa zonke i-prompts yakho kwi-externally kuphela ukuze ukunciphise ukuhlanganiswa kunye ne-evaluation tool yeentlobo ze-ML.




Ukuba ufumane le post, nceda uqhagamshelane nqakraza ezininzi.